Google Launches Gemini Decisions API for 150ms Single-Pass Multimodal Routing at $0.10 per 1M Tokens
Powered by the new Gemini Decision 1 model in Google AI Studio and Vercel AI Gateway, the endpoint evaluates up to 1,000 boolean, multiple-choice, and rubric questions in one forward pass with zero output tokens billed.
At a Glance
- What changed: Google launched the Gemini Decisions API powered by Gemini Decision 1 (gemini-decision-1), scoring up to 1,000 closed-set questions in a single forward pass across text, images, audio, video, and PDFs (up to 256K tokens).
- Who can use it: Developers using Google AI Studio (v1beta API), Vercel AI Gateway (model ID google/gemini-decision-1), OpenRouter, LangChain, LiteLLM, Pydantic AI, Pipecat, LiveKit, and OpenCode.
- What it costs: $0.10 per 1M input tokens with $0.00 (zero) output tokens billed—nearly 5x cheaper across Google’s shared benchmark suite than Gemini 3.5 Flash-Lite on the Interactions API.
- Why should I care: Agent tool routing, content moderation, and rubric grading no longer require slow, expensive autoregressive JSON generation; you get calibrated 0-to-1 confidence scores in 150–250 ms end-to-end.
Google DeepMind has launched the Gemini Decisions API, a purpose-built developer endpoint powered by the new Gemini Decision 1 model (gemini-decision-1) in Google AI Studio and Vercel AI Gateway (google/gemini-decision-1). Instead of generating text token by token, the endpoint evaluates up to 1,000 independent closed-set questions against a shared multimodal input in a single forward pass—returning calibrated probabilities in 150 to 250 milliseconds at $0.10 per 1 million input tokens and $0.00 for output tokens.
The release arrives alongside Google Cloud’s new Professional Agentic Architect certification, underscoring a broader architectural shift in late 2026: separating fast, deterministic agent routing and policy gating from heavier reasoning and code-generation calls.
How Single-Pass Decision Scoring Replaces Token Decoding
In standard LLM workflows, asking a model to classify an image, select one of 20 agent tools, or grade a document against a 5-point rubric requires autoregressive decoding—generating JSON keys and string values one token at a time. Even with small models, sequential token generation introduces tail latency, schema-parsing failures, and uncalibrated confidence estimates.
Written by Google DeepMind Product Manager Min Choi and Principal Scientist Felix Yu, the technical specification for Gemini Decision 1 bypasses text generation altogether. Developers pass a single shared multimodal context—combining text, images, audio, video, and PDF documents up to 256,000 tokens—and attach a dictionary of up to 1,000 independent questions. The model computes normalized probability distributions across all questions in a single forward pass:
- Median Model Compute Latency: 50 to 100 ms (150 to 250 ms end-to-end across multimodal inputs up to 16K tokens, remaining nearly flat from 1 to 100 questions per request).
- Zero Output-Token Billing: Because the model returns forward-pass probability scores rather than decoded text tokens, Google bills strictly on input volume at $0.10 per 1M input tokens. Across Google’s shared routing and ranking benchmark suite, that pricing structure makes Gemini Decision 1 nearly 5x cheaper than
Gemini 3.5 Flash-Liteon the standard Interactions API. - Calibrated Confidence Scores (0 to 1): Every response includes a calibrated probability score so application code can automatically route low-confidence edge cases to a deeper reasoning model (such as Gemini 4 Carbon) or a human reviewer.
The Three Decision Primitives & Platform Comparison
A single call to client.decisions.create (in Google’s google-genai SDK) or experimental_decide (in Vercel’s AI SDK) can mix three distinct question types against the same multimodal payload. However, cross-referencing Google’s documentation with Vercel’s AI Gateway changelog reveals an important interface distinction for teams processing PDFs, audio, or video:
| Capability / Dimension | Google AI Studio (Gemini Decisions API) | Vercel AI Gateway (AI SDK 7) | Vercel TypeSafe & Decisions REST API |
|---|---|---|---|
| Model Identifier | gemini-decision-1 |
google/gemini-decision-1 |
google/gemini-decision-1 |
| Supported Input Modalities | Text, Images, Audio, Video, PDFs (up to 256K tokens) | Text, Images, Audio, Video, PDFs | Text and Images only |
Yes/No Primitive (noul / boolean) |
Probability (0–1) with optional true/false criteria | type: 'boolean' via experimental_decide |
Supported (Boolean probability) |
Multiple Choice (choice) |
2 to 255 candidate options with normalized probabilities | Supported (Choice winner + distribution) | Supported (Choice winner + distribution) |
Ordered Rubric (score) |
Ordered scale starting at 0; returns continuous weighted score | Supported (Rubric weighted score + per-level probabilities) | Supported (Rubric weighted score) |
| Pricing & Output Billing | $0.10 / 1M input tokens; $0.00 output tokens | Unified AI Gateway budget tracking; $0 output tokens | Unified AI Gateway budget tracking; $0 output tokens |
Ecosystem Rollout and Enterprise Agent Architecture
Beyond Google AI Studio and Vercel AI Gateway, day-one integrations for the Gemini Decisions API are rolling out across OpenRouter, LangChain, LiteLLM, Pydantic AI, Pipecat, LiveKit, and OpenCode. That broad day-one coverage—particularly in real-time voice and video frameworks like Pipecat and LiveKit—targets low-latency turn detection, barge-in classification, and live safety guardrails where a 150ms budget is mandatory.
To formalize those multi-model routing patterns for enterprise engineering teams, Google Cloud simultaneously introduced its Professional Agentic Architect certification. The credential combines a proctored architectural exam with hands-on Google Skills labs focused on multi-step, multi-application agent orchestration, and includes 35 free monthly learning credits through the Gemini Agent Ready program.
What Engineering Teams Should Watch
For teams currently spending thousands of dollars per month running moderation checklists, intent classifiers, or RAG chunk relevance grading through general-purpose chat completions, moving bounded checks to gemini-decision-1 offers an immediate 80% cost reduction and eliminates JSON formatting retries. Teams running on Google Cloud Vertex AI should note that while Google AI Studio and third-party gateways are live today, native Vertex AI integration is still listed as "coming soon."
Why It Matters
Most production AI pipelines waste latency and output-token budget using full generative models just to make bounded choices—such as routing a support ticket, picking an agent tool, or checking a compliance rule. By scoring candidate options directly in a single forward pass and eliminating output-token billing entirely, Google’s Gemini Decisions API turns multimodal classification and routing into a sub-250ms, fixed-input-cost primitive.
What Is Confirmed
Confirmed via primary documentation from Google DeepMind (Min Choi and Felix Yu), Vercel AI Gateway, and Google Cloud: Gemini Decision 1 processes up to 256K multimodal input tokens and up to 1,000 independent questions per request; median model compute is 50–100 ms (150–250 ms end-to-end up to 16K context); pricing is $0.10 per 1M input tokens with zero output tokens billed; and Vercel AI Gateway supports full PDF/audio/video multimodal inputs via AI SDK 7 while its TypeSafe and Decisions REST endpoints currently accept text and images.
What Remains Unclear
Google states that native support for Gemini in Google Cloud (Vertex AI) is "coming soon" without specifying a general-availability date, and enterprise SLA tiers for high-concurrency batch routing have not yet been published.
Sources & Underlying Evidence (3)
We link directly to the primary documents supporting our reporting and distinguish company claims from independently checked facts.
- Gemini Decisions API: Fast, calibrated multimodal decisionsPublisher: Google Official Blog (DeepMind) • Type: Official AnnouncementSource date: Nov 11, 2026
- Gemini Decision 1 now available on AI GatewayPublisher: Vercel Changelog • Type: Product ChangelogSource date: Nov 11, 2026
- Get certified as a Google Cloud Professional Agentic ArchitectPublisher: Google Cloud Official Blog • Type: Official AnnouncementSource date: Nov 11, 2026
